Causal Semantics Refinement Model for Event Causality Identification via information bottleneck representation learning

Identifying the causal relationship between events is an important foundation for understanding text semantics and supporting downstream reasoning tasks. However, existing methods are often hindered by semantic noise and redundant features when dealing with implicit causal relationships, making it difficult to capture causal signals that lack explicit labels effectively. To solve this problem, this paper proposes a Causal Semantics Refinement Model (CaSR) based on information bottleneck optimization. This method first constructs semantic-level positive and negative event pairs to guide the model to align with potential causal semantics and establish comparable references. Secondly, non-causal perturbation and the introduction of semantic restoration mechanisms force the model to identify causal core features during denoising, thereby enhancing the robustness to non-explicit causal relationships. Finally, combined with the information bottleneck optimization strategy, the key information is retained while compressing the redundant features, thereby enhancing the model’s representation ability of implicit causal relationships. The experimental results on the mainstream EventStoryLine, Causal-TimeBank, and the Uyghur News Datasets constructed in this study show that CaSR achieves the current optimal performance in the event causality identification task, with the F1 score increasing by 0.7%-2.4%. The Uyghur News Datasets are available at https://github.com/jinghaipeng/Causal-Semantics-Refinement-Model.